Playbook

Why the same work gets repeated #

A release or recovery may succeed after several attempts, but the working procedure remains buried in a transcript. The next agent repeats the search, including the mistakes.

Playbook starts by reading the commands and instructions already responsible for that work. It repairs an incomplete procedure, removes a stale one, or captures the missing steps where the next agent will find them. A safe one-off can end without adding anything to the repository.

Bad instructions are work too #

A Pixel Brite pagination skill existed even though pagination was prohibited. The skill was deleted; the actual invariant stayed in the project instructions. Large Bluetooth and Iroh runbooks had also accumulated in the top-level agent file. Those moved behind narrow triggers, with detailed evidence in ordinary project documentation.

Ordinary tasks load fewer instructions. Bluetooth and Iroh tasks still get the detailed procedures when they need them.

Stringproof became a CLI #

Apple .strings files can still parse after translation while losing a key or placeholder, gaining prompt residue, or mixing scripts. Those checks follow exact rules, so Playbook chose a CLI instead of another skill.

The result is stringproof check: one command reads a directory, reports stable findings, and returns a distinct exit code for clean input, translation damage, or an invocation failure.

Put the work where it fits #

Existing command or procedure

Reuse or repair it

Stale or duplicate instructions

Remove them

Repeatable steps

Script or CLI

Recurring decisions

Skill

External service or data source

Integration

Rule for every repository task

AGENTS.md

Safe one-off

Nothing

Use what you leave behind #

The agent finishes the current task through the command or instructions it keeps. Real inputs expose missing arguments, assumptions, and recovery steps. The procedure is ready when completing the job no longer depends on remembering the conversation.